Grok 4.6 is cheap — until your agent crosses 200K tokens

SpaceXAI (the company formerly called xAI) shipped Grok 4.6 on 12 August, built explicitly for long-running agents — multi-step research, whole-codebase work, rough idea to shipped v1. It lands at $2/M input, $6/M output with a 500K-token context window, and scores 61 on the Artificial Analysis Intelligence Index — a statistical tie with GPT-5.6 Sol Max, one point behind Fable 5 Max at 62. Claude Opus 5 still tops the index.

Here's the part the hype posts skip: cross 200K prompt tokens and xAI rebills the entire request at $4/$12 per million. That's precisely the workload Grok 4.6 is marketed for — agent loops accumulate context by design. Against GPT-5.6 Sol's $5/$30, the real saving is closer to 60–80%, and only below the threshold. A faster variant costs double again, with no separate model ID.

On code it's improved but not leading: DeepSWE v1.1 at 65.9% (up 11.9 points generationally, still behind Sol Max's 73%) and Terminal-Bench v3.0 at 26% — nearly double Grok 4.5, yet last among the four models compared. Its genuine win is knowledge work: 1753 Elo on GDPval-AA v2, top of the field. Available now via the xAI API, Cursor, Grok Build, OpenRouter, Vercel and Cloudflare. No open weights.

Why it matters: frontier intelligence is converging on parity, so the fight has moved to price-per-token — and to whose fine print you read before wiring up an agent.